VR-Recon: Visual Refiner for Fine-Grained Human Reconstruction

To reconstruct 3D human figures with fine-grained details from sparse views, this paper proposes a novel framework called VR-Recon, which is based on a viewpoint refiner. In this framework, we first decouple geometric features from multi-scale feature learning by integrating an MSG image encoder, aiming to extract high-quality multi-scale geometric features. Next, we employ a viewpoint refiner to accurately capture critical fine-grained geometric information in each independent view space, thereby avoiding the weighted average fusion of features and eliminating the influence of invalid information. This method is trained using a consistent cosine loss and a cascaded view-space geometric loss. The former encourages the MSG to accurately perceive fine geometric changes from the image, while the latter helps the viewpoint refiner discern critical information in different view spaces. Experimental results demonstrate that, compared to state-of-the-art methods, VR-Recon exhibits robust performance in capturing fine geometric details.